Skip to main content
Itoflow’s research agent can work in your GitHub repositories and call tools on your own MCP servers, beside Itoflow’s engine and data. Your models stay in your code, your tools stay on your servers, and the agent brings the data, the tests and the review. What you’ll get: research that runs your own code or calls your own tools, and, when you ask, a draft pull request in your repository with the change and its results.

Before you start

Try it

With Itoflow connected to your assistant, paste this into your assistant:

Open in Claude

Start this in Claude with the Itoflow plugin.

Open in ChatGPT

Start this in ChatGPT with the Itoflow plugin.
Using another client? Copy the prompt with the button on the code block.Replace OWNER/REPO with your repository and PATH with the backtest’s path in it.
To keep code and data on your own machine instead, use the local worker: Itoflow plans the analysis and your local Codex or Claude Code session runs each step.

What you get back

The finished turn holds the code the agent ran, its output and the files it changed, with a link to the pull request. Asked for a draft, the agent opens one. Your branch protection still applies. With your own MCP server, the agent finds the tools it needs in the server’s catalog, calls them, and uses what they return in its research. Analyze your Interactive Brokers portfolio walks through one: IBKR’s own server reads your positions, and Itoflow’s risk model measures them.

How it’s kept safe

How we connected customers’ MCP servers and GitHub repos to our research agent, October 7, 2026, explains the design:
  • Tokens stay out. The model and the sandbox never see a token. Credentials are added outside the sandbox after Itoflow checks you own the connection.
  • Narrow GitHub access. The GitHub app asks only for code, pull requests and CI results: no workflows, admin rights or secrets.
  • Outside text is data. Text from a repository or a tool counts as untrusted input, not as instructions.
  • Merges only when you ask. The agent merges a pull request only when you ask it to merge that one, and GitHub’s required checks still apply.
  • No blind repeats. A write whose outcome is unknown is reported as unknown and never sent twice.
  • Tools stay out of the way. A server’s tools sit in a catalog the agent searches, so a server with dozens of tools doesn’t crowd the model’s context.

Limits

  • An MCP server needs a public HTTPS URL and Streamable HTTP. It signs in with OAuth or a bearer token.
  • If a server’s tool catalog can’t be read, its tools may be missing from a turn. Reconnect when the connection says sign-in is required.

Next steps